[Lens] standardize references between columns

[drewdaemon]

Background
Today, referenced columns in the indexpattern datasource are always intermediary columns that are used "under the hood." If we conceptualize the columns as a graph with the columns being the nodes and the references being the edges, they form a series of DAG trees, the top level columns that are directly used in the visualization being the roots.

No references exist between trees except for one: sort-by in the top values function. In the code, this reference is not modeled in the same way as the others.

We could standardize how references are represented in the code to provide an easy and consistent way to understand and change which columns reference each other. One option is to follow the pattern of saved object references and store all column references in a list.

Describe a specific use case for the feature:
If we standardized the way all references between columns are represented we would unlock

• Easier ways to detect and update sort-by in the top values function. This would be really useful in aggregation optimization tasks such as [Lens] Optimize redundant formula function aggregations #135265 and [Lens] Optimize percentiles fetching #126941
• Formula references (e.g. one function uses the output of another — @flash1293 feel free to augment/correct this explanation)

[elasticmachine]

Pinging @elastic/kibana-vis-editors @elastic/kibana-vis-editors-external (Team:VisEditors)